Multi-Level Based Pedestrian Attribute Recognition

Pedestrian attribute recognition is a challenging task in computer vision due to its multi-label nature. Many algorithms have been proposed to solve this problem, but a better one is still needed. In this paper, we propose an effective and novel method. Given a multi-label pedestrian image, our Multi-level Aggregate Network (MAN) generates feature maps at three different levels and aggregates three predictions as final output. The proposed network is trained in an end-to-end manner with only image-level annotations. Extensive experiments are performed on the two largest pedestrian attribute datasets i.e. the PETA dataset and PA-100K dataset. We achieve state-of-the-art results without other information.